A well logging lithology prediction method based on multi-well incremental learning

Abstract

The present application relates to the technical field of oil and gas exploration, and discloses a well logging lithology prediction method based on multi-well incremental learning. The construction and training process of the prediction model comprises: constructing a global incremental data set with single drilling as an incremental unit, and preprocessing and standardizing the well logging data of each drilling; extracting the well logging time sequence features of each well logging data sample in the global incremental data set along the time branch and the channel branch respectively, and obtaining the global time sequence representation by fusion; decoupling the global time sequence representation into lithology causal features and domain-specific pseudo features using a mask; performing intra-class causal disturbance and inter-class causal disturbance on the decoupled features respectively, and constructing a global joint loss function; incrementally iteratively training the historical prediction model parameters in order with single drilling as an incremental unit, saving the current prediction model parameters, and sequentially completing multi-well incremental learning. The present application can improve the lithology recognition accuracy of new drilling and the model cross-well generalization ability.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration technology, specifically to a logging lithology prediction method based on multi-well incremental learning. Background Technology

[0002] In oil and gas exploration and development operations, well logging data is the core foundational data for determining formation lithology, identifying reservoir locations, and evaluating oil and gas reserves. Accurate and efficient lithology identification results can effectively guide drilling trajectory optimization and reduce exploration and development costs. Currently, the industry generally relies on human experience combined with statistical algorithms to complete lithology interpretation. This model is highly dependent on the professional experience of geological engineers, has a large subjective error, and is difficult to adapt to the batch analysis of data from multiple wells in large-scale work areas. This results in low work efficiency and fails to meet the operational requirements of modern, efficient oil and gas exploration.

[0003] With the gradual popularization of deep learning technology, many intelligent lithology prediction methods based on temporal neural networks and feature classification networks have been widely applied. These methods can automatically uncover the potential correlation between well logging curves and formation lithology, significantly reducing the cost of manual interpretation. However, most existing intelligent prediction models are statically trained, with fixed parameters after training, and are only suitable for drilling data within a single exploration block. When new unknown wells are added to the work area or there are shifts in the geological distribution between wells, the model is prone to problems such as declining prediction accuracy and insufficient generalization ability, making it difficult to adapt to the actual operational scenarios of multi-well progressive exploration.

[0004] In addition, existing well logging lithology prediction models generally suffer from defects such as single feature extraction and poor feature decoupling. The models cannot distinguish between the inherent lithological characteristics of the well and external interference features such as instrument noise and block environment, and are easily affected by invalid and redundant information. At the same time, conventional incremental learning algorithms are prone to catastrophic forgetting problems when adapting to multi-well scenarios. While learning the patterns of newly added well data, they lose the lithological knowledge accumulated from historical wells. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides a logging lithology prediction method based on multi-well incremental learning. Relying on dual-channel encoding, orthogonal feature decoupling, and dual-causal intervention mechanisms, it completes multi-well incremental training without storing historical raw data, thereby improving the accuracy of lithology identification for newly drilled wells and the model's cross-well generalization ability.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A well logging lithology prediction method based on multi-well incremental learning, the construction and training process of the prediction model includes: Construct a global incremental dataset with each well as an incremental unit, and preprocess and standardize the logging data of each well. A dual-channel coding structure is constructed, and the logging time-series features of each logging data sample in the global incremental dataset are extracted along the time branch and the channel branch respectively. Based on the formation data fluctuation state, the dual-branch features are adaptively fused to obtain the global time-series representation. Based on global temporal representation, causal feature masks and pseudo-feature masks are generated through a dual-branch independent mapping network. An F-norm-based mask orthogonal regularization loss is introduced to make the two types of masks orthogonal. The masks are used to decouple the global temporal representation into lithological causal features and domain-specific pseudo-features. Intra-class causal perturbation and inter-class causal perturbation are performed on the decoupled lithological causal features and domain-specific pseudo-features, respectively. Intra-class causal intervention loss and inter-class causal intervention loss are constructed based on the lithological classifier. The intra-class causal intervention loss, inter-class causal intervention loss and mask orthogonal regularization loss are fused based on dynamic adaptive weights to construct a global joint loss function. Using a single well as an incremental unit, historical prediction model parameters are loaded sequentially for incremental iterative training, and the current prediction model parameters are saved. This process is repeated for multiple wells to complete incremental learning.

[0007] In one embodiment, the construction of a global incremental dataset with a single well as the incremental unit specifically includes: Each well within the exploration area is divided into an independent incremental domain, and a global incremental dataset is generated. ,in, , This represents the total number of wells drilled. For the first The subset of data corresponding to the wellhead drilling. For the first The first drilling in the well well logging data sample, For the number of logging data channels, for lithological labels For the first Total number of logging data samples in wells; It is the set of real numbers.

[0008] In one embodiment, the construction of the dual-channel encoding structure, which extracts the logging time-series features of each logging data sample in the global incremental dataset along the time branch and the channel branch respectively, specifically includes: The dual-channel encoding structure includes a time encoder and a channel encoder; well logging data samples Perform sliding window segmentation to obtain time series samples. The time encoder uses time-series samples For input, a two-layer one-dimensional convolutional layer is used to extract local temporal correlation features along the well depth direction layer by layer. After batch normalization, ReLU activation, and average pooling, the temporal evolution representation is output. The channel encoder takes the transposed temporal sample as input, uses a two-layer one-dimensional convolutional layer to complete feature interaction in the channel dimension, and after batch normalization and ReLU activation, outputs a multi-parameter coupled representation. Modeling the intrinsic physical relationships between multiple parameters: ; ; The temporal evolution representation of the output of the time branch; Multi-parameter coupling characterization of channel branch output; This is a matrix transpose operation; These are the learnable network parameters corresponding to the time encoder and the channel encoder, respectively; the logging time series features include time series evolution representation and multi-parameter coupling representation.

[0009] In one embodiment, the adaptive fusion of dual-branch features based on formation data fluctuation states to obtain a global time series representation specifically includes: Dynamically calculate fusion weights based on the temporal variance of a single batch of samples: ; For adaptive fusion weights; For activation functions; This is the variance calculation function; Representing the Inside the well, the first Time-series data corresponding to each logging data channel; Number the wells; This represents the total number of well logging data channels. By fusing features from the time branch and channel branch outputs using adaptive fusion weights, an enhanced global temporal representation is obtained. And connected to the lithology classifier Preliminary lithological prediction completed: ; ; The temporal evolution representation of the output of the time branch; This represents the multi-parameter coupling of the channel branch output. For the first All learnable network weights and bias parameters corresponding to the lithology classifier for each well; This is for preliminary prediction of lithological results.

[0010] In one embodiment, the step of generating causal feature masks and pseudo-feature masks based on global temporal representation through a dual-branch independent mapping network, and introducing an F-norm-based mask orthogonal regularization loss to make the two types of masks orthogonal, specifically includes: ; ; These are two sets of single-layer fully connected neural networks with identical structures and independent parameters. As a causal feature mask; This is a pseudo-feature mask; The normalized activation function; Introducing an F-norm-based mask orthogonal regularization loss Quantify the overlap between causal feature masks and pseudo-feature masks, by Backpropagation minimizes the matrix inner product, achieving explicit orthogonality constraints: ; This represents the matrix transpose operation; It is the F-norm.

[0011] In one embodiment, the decoupling of the global temporal representation into lithological causal features and domain-specific pseudo-features using a mask specifically includes: It employs element-wise multiplication and utilizes orthogonal causal feature masks. and pseudo-feature mask Global temporal representation Feature decoupling and separation were performed to isolate lithological causal characteristics. Domain-specific pseudo-features : ; ; This is an element-wise matrix multiplication operation.

[0012] In one embodiment, the step of performing intra-class causal perturbation and inter-class causal perturbation on the decoupled lithological causal features and domain-specific pseudo-features, respectively, specifically includes: Intraclass causal perturbation: For any target logging data sample, other logging data samples with the same lithology label as the target logging data sample are selected from the global incremental dataset. The domain-specific pseudo-features of the selected logging data samples are extracted and replaced with the domain-specific pseudo-features of the target logging data sample, while retaining the lithological causal features of the target logging data sample itself. Superimposed lithological causal characteristics With the updated domain-specific pseudo-features Generate intra-class perturbation samples : ; An index for the target well logging data sample; An index representing other well log samples whose lithology labels match the target well log sample; Inter-class causal perturbation: For any target well logging data sample, and with the introduction of lithological causal features of well logging data samples with other lithological labels, the domain-specific pseudo-features of the target well logging data sample are replaced, while preserving the lithological causal features of the target well logging data sample itself. Superimposed lithological causal characteristics Based on the causal characteristics of exogenous lithology, inter-class perturbation samples are generated. : ; An index representing a logging data sample whose lithology label differs from the target logging data sample. Lithological causal characteristics of logging data samples that differ from the lithological label of the target logging data sample.

[0013] In one embodiment, the construction of a global joint loss function based on dynamically adaptive weight fusion of intra-class causal intervention loss, inter-class causal intervention loss, and mask orthogonality regularization loss specifically includes: Calculate intraclass causal intervention loss : ; For the i-th target logging data sample, there is an intra-class disturbance sample; The cross-entropy loss function; The true lithology label for the i-th target logging data sample; As a lithology classifier, These are all the learnable network weights and bias parameters corresponding to the lithology classifier; Calculate inter-class causal intervention loss : ; For the inter-class perturbation sample of the i-th target logging data sample; Intra-class causal intervention loss based on dynamic weight fusion Inter-class causal intervention loss Mask orthogonal regularization loss The global joint loss function is obtained. : ; The mask orthogonal regularization loss balancing hyperparameter, It is a dynamic adaptive weight, calculated based on KL divergence, used to quantify the distribution offset between the current drilling lithological causal characteristics and the historical drilling lithological causal characteristics.

[0014] In one embodiment, the dynamic adaptive weights are calculated as follows: ; ; The well-to-well domain distribution divergence; Let KL divergence function be used. The characteristic probability distribution function; The lithological causal characteristics of newly drilled wells; The cumulative lithological causal characteristics of historically trained wells; This refers to the divergence scaling hyperparameter; For dynamic adaptive weights, This is the activation function.

[0015] In one embodiment, the step of using a single well as an incremental unit, sequentially loading historical prediction model parameters for incremental iterative training, and saving the current prediction model parameters to complete multi-well incremental learning sequentially includes: During the first incremental domain training, the prediction model parameters are randomly initialized, the prediction model parameters are updated based on the global joint loss function, and the optimal prediction model is saved. During subsequent incremental domain training, the optimal prediction model parameters from the previous incremental domain are loaded as initial parameters, and the prediction model parameters are iteratively updated based on the current incremental domain data until the prediction model converges and the optimal model parameters of the current incremental domain are saved. During training, historical drilling and logging data samples are not cached; only the prediction model parameters are saved. The prediction model parameters include those of the dual-channel coding structure, the bi-branch independent mapping network, and the lithology classifier.

[0016] Compared with the prior art, the beneficial technical effects of the present invention are: 1. Enhance the comprehensiveness and adaptability of feature extraction. This invention constructs a dual-channel differentiated coding structure to mine well logging data features from both the time-series and channel dimensions, taking into account both the vertical formation evolution law of a single well and the petrophysical coupling relationship of multiple well logging parameters. At the same time, it configures adaptive fusion weights to dynamically adjust the proportion of dual-branch features according to the fluctuation state of formation data, effectively solving the problem of the single dimension of feature extraction in traditional time-series coding, and significantly improving the lithological characterization ability of thin reservoirs and complex heterogeneous formations.

[0017] 2. Optimize feature decoupling quality and reduce external operating condition interference. This invention designs a dual-branch orthogonal mask generation mechanism, using independent mapping networks to generate causal feature masks and pseudo-feature masks respectively, and adds explicit mask orthogonal regularization loss to achieve complete decoupling of the two types of features from a mathematical perspective, accurately separating the inherent lithological features of the formation from drilling-specific interference features; combined with intra-class and inter-class bidirectional causal perturbation strategies, it forces the model to complete classification and discrimination based on the essential lithological features, weakening the negative impact of irrelevant factors such as well environment and equipment noise.

[0018] 3. Adaptable to multi-well incremental exploration scenarios, balancing anti-forgetting capabilities and engineering practicality. This invention uses a single well as an incremental unit to build an incremental learning framework, combined with a dynamic adaptive weight joint loss function, to adaptively adapt to data distribution shifts between different wells. The entire training process does not require caching historical drilling samples; it only iteratively saves network parameters, resulting in low memory usage. This solves the shortcomings of traditional static models in adapting to new wells and alleviates the catastrophic forgetting problem during incremental training, enabling efficient service for large-scale intelligent lithology prediction operations in oil and gas fields. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] like Figure 1 As shown, this invention provides a well logging lithology prediction method based on multi-well incremental learning. The prediction model construction and training process includes the following steps: S1. Construct a global incremental dataset with each well as an incremental unit, and preprocess and standardize the logging data of each well. S2, construct a dual-channel coding structure, extract well logging time series features of each well logging data sample in the global incremental dataset along the time branch and channel branch respectively, and adaptively fuse the dual-branch features based on the formation data fluctuation state to obtain a global time series representation; S3, based on global temporal representation, generates causal feature masks and pseudo-feature masks through a dual-branch independent mapping network, and introduces an F-norm-based mask orthogonal regularization loss to make the two types of masks orthogonal. The masks are used to decouple the global temporal representation into lithological causal features and domain-specific pseudo-features. S4. Intra-class causal perturbation and inter-class causal perturbation are performed on the decoupled lithological causal features and domain-specific pseudo-features, respectively. Intra-class causal intervention loss and inter-class causal intervention loss are constructed based on the lithological classifier. Intra-class causal intervention loss, inter-class causal intervention loss and mask orthogonal regularization loss are fused based on dynamic adaptive weights to construct a global joint loss function. S5 uses a single well as an incremental unit, loads historical prediction model parameters sequentially for incremental iterative training, saves the current prediction model parameters, and completes multi-well incremental learning in sequence.

[0022] The present invention will be described in detail below in several parts.

[0023] 1. Construction and preprocessing of multi-well incremental datasets.

[0024] (1) Define the domain incremental task and time series samples: This invention proposes a well-level incremental task construction method adapted to multi-well oil and gas exploration. It uses a single well as the basic incremental unit, divides different wells in the exploration area into independent incremental domains, and constructs an incremental domain sequence in an orderly manner to complete incremental learning and training in multi-well scenarios.

[0025] , ; In the formula, This is a global incremental dataset for the entire exploration area; The total number of wells participating in model training within the work area; For the first Subdataset corresponding to the wellhead drilling; For the first The first drilling in the well A sample of raw well logging data; The number of logging data channels covers four mainstream logging parameters: gamma rays, shallow and deep resistivity, sonic transit time, and density. Artificially labeled lithological tags for corresponding time series samples; For the first The total number of samples within each well. In actual exploration scenarios, the model sequentially learns the data distribution and lithological patterns of multiple historical wells. When a new, unknown well is added to the work area, this new well is treated as a completely new incremental domain, and intelligent lithological prediction is completed based on the learned historical well knowledge.

[0026] (2) Data cleaning and standardization: To address the outlier data generated by instrument noise, mud environment interference, and complex formation heterogeneity during drilling operations, outlier removal and high-frequency noise filtering preprocessing are performed sequentially on the original multi-channel time-series logging data. Then, the extreme value normalization method is used to standardize the data across the entire domain, constraining all logging channel data to a fixed numerical range, eliminating data amplitude deviations caused by different drilling, different instruments and equipment, and different formation environments, and outputting a standardized multi-well time-series sample set as a fixed input for subsequent prediction models.

[0027] (3) Dataset partitioning: Standardized datasets corresponding to each incremental domain are processed according to a fixed ratio. The model is split into training, validation, and test sets. The training set is used for iterative updates of model parameters, the validation set is used for hyperparameter optimization and early stopping of the model to prevent overfitting, and the test set is used for final quantitative evaluation of model performance.

[0028] 2. Decoupling of temporal feature encoding and orthogonal features.

[0029] (1) Temporal feature encoding: Conventional time-series coding structures can only extract shallow, single-dimensional time-series features, making it difficult to simultaneously consider both the vertical stratigraphic evolution patterns of a single well and the petrophysical correlation features between multiple logging parameters. To address this, this invention proposes a dual-channel differentiated coding structure for multi-well incremental prediction. It employs a dual-branch parallel coding mode to separately extract time-series evolution features along the well depth direction and the coupling features of multiple logging parameters. Adaptive weighting is used to fuse the dual-branch features, effectively improving the model's lithological generalization prediction capability for newly added unknown wells.

[0030] Two-branch feature independent encoding: the original sample is This represents a single multi-channel logging data line without a time window; to adapt to time series model training, [the following is used]... Perform sliding window segmentation to obtain time-series samples of a single input to the model. ;in The time-series sliding window length, The number of logging channels is defined. A time-coding branch and a channel-coding branch are constructed separately, independent of each other and complementary in function: the time branch uses the original sample as input to capture dynamic changes in lithology along the vertical well depth dimension; the channel branch transposes the sample matrix to model the intrinsic physical relationships between multiple parameters such as gamma, resistivity, and acoustic waves. The calculation formulas are as follows: ; ; Explanation of parameters in the formula: The temporal evolution representation of the output of the time branch; Multi-parameter coupling characterization of channel branch output; This is a matrix transpose operation; These are the learnable network parameters for the time encoder and the channel encoder, respectively.

[0031] This invention defines specific implementation structures for two types of encoders, ensuring the integrity and reproducibility of the technical solutions. Both the time encoder and the channel encoder employ a two-layer one-dimensional convolutional structure, adaptable to well logging time-series data feature extraction tasks. Time encoder Two one-dimensional convolutional layers with a kernel size of 3 are set up to extract local temporal correlation features in the well depth direction layer by layer. After each convolutional layer, a BN normalization layer and a ReLU activation function are configured. Finally, an average pooling layer is connected to compress the feature dimension, focusing on characterizing the continuous vertical formation variation law of a single well.

[0032] Channel encoder The input is the transposed channel dimension data. It also adopts a two-layer one-dimensional convolutional structure with a fixed kernel size of 1. It completes feature interaction only in the channel dimension. It is paired with a BN normalization layer and a ReLU activation function, which is specifically used to model the coupling relationship between multiple types of well logging curves and suppress invalid channel redundancy information.

[0033] Adaptive fusion weight calculation: To avoid the limitation of fixed fusion weights failing to adapt to different formation conditions, the fusion coefficients are dynamically calculated based on the temporal variance of a single batch of samples. When the formation exhibits significant fluctuations, the weight of the time branch is increased; when the formation is smooth and homogeneous, the channel coupling characteristics are emphasized. The calculation formula is as follows: ; Explanation of parameters in the formula: Adaptive fusion weights for dual-branch features; This is a general activation function used to constrain values ​​to the 0-1 range; This is a variance calculation function used to measure the degree of fluctuation in formation data; Representing the Inside the well, the first Time series data corresponding to each logging channel; Number the wells; This represents the total number of logging channels.

[0034] Global Feature Fusion and Lithology Prediction: Based on adaptive weighted fusion of bi-branch features, the final enhanced well logging time series characterization is obtained. And connect it to the lithology classifier to complete the preliminary lithology prediction: ; ; Explanation of parameters in the formula: This is the enhanced global temporal representation after fusion; In a preferred embodiment, the lithology classifier can adopt the following structure: the classifier adopts a three-layer fully connected neural network, which consists of an input mapping layer, a feature transition layer, and an output classification layer. The first layer performs dimensional mapping on the high-dimensional fused features, the middle layer uses the ReLU activation function to complete the nonlinear feature transformation, and the last layer is configured with the Softmax activation function to output the predicted probability distribution corresponding to each type of lithology. For the first All learnable network weights and bias parameters corresponding to the lithology classifier under each incremental domain; This is the preliminary predicted lithology result output by the model. This encoding method overcomes the limitations of general time series models, taking into account both the vertical evolution law of strata and the multi-parameter physical coupling relationship, and significantly improves the identification accuracy of thin reservoirs and abrupt strata.

[0035] (2) Generation of orthogonal mask matrix: To completely separate inherent formation lithological features from drilling-specific interference features in well logging data and improve feature decoupling quality, this invention proposes a bi-branch adaptive explicit orthogonal mask generation mechanism. Two independent mapping networks are set up to generate corresponding mask matrices, and explicit orthogonal constraints are added to constrain the distribution of the two types of masks at the matrix operation level. This, combined with global temporal representation, completes the differential mask solution. The specific calculation formula is as follows: ; ; Explanation of parameters in the formula: Two sets of single-layer fully connected neural networks with identical structures and independent parameters are used to generate causal feature masks and pseudo-feature masks, respectively. This is a causal feature mask matrix used to filter lithological essence discrimination features; This is a pseudo-feature mask matrix used to locate interference features of blocks, devices, and operating conditions; This is a normalized activation function used to constrain the mask values ​​to the range of 0 to 1.

[0036] To further eliminate mask overlap defects and ensure feature decoupling accuracy, an F-norm orthogonal constraint loss is introduced to quantify the overlap between two mask matrices. The matrix inner product is minimized through backpropagation of the loss, achieving explicit orthogonality constraints. The orthogonal loss formula is as follows: ; Explanation of parameters in the formula: For mask orthogonal regularization loss; This represents the matrix transpose operation; The F-norm is a well-known matrix operation method. The loss aims to optimize the matrix inner product to approach 0, ensuring that there is no overlap between the causal feature mask and the pseudo-feature mask, thus avoiding the problem of cross-mixing of the two types of features from the root.

[0037] (3) Decoupling and splitting of dual-type features: The global time series is represented by two sets of orthogonal mask matrices using an element-wise multiplication operation. Feature decoupling and decomposition are performed to separate causal features and spurious correlation features from the original representation. The calculation formula is as follows: ; ; Explanation of parameters in the formula: This is an element-wise matrix multiplication operation. The lithological causal characteristics that are stable across blocks and working conditions are the core inherent characteristics that distinguish different stratigraphic lithologies; Domain-specific pseudo-correlation features include invalid interference information such as drilling block differences, instrument noise, and operating conditions; the decoupled results... The data is output synchronously and used as input data for the next stage of the dual-causal intervention module.

[0038] 3. Dual-causal intervention constraints and domain incremental model iterative optimization.

[0039] (1) Calculation of intraclass causal perturbation and intraclass intervention loss: To mitigate data distribution discrepancies for the same lithology across different wells and enhance the cross-well stability of causal features, this invention employs an intra-class causal perturbation strategy. The core idea of ​​this strategy is to fix the unique lithological causal features of the current sample and replace its pseudo-features containing interfering information, thereby eliminating the negative impact of pseudo-features on lithology identification. The specific operation is as follows: For any target sample, other samples with completely identical lithological labels are selected from the dataset. Pseudo-features of these dissimilar but homologous samples are extracted and replaced with the original pseudo-features of the target sample, retaining only the target sample's own causal features. The corresponding feature update rules are as follows: ; In the formula: This represents the target sample currently awaiting processing. This represents a homologous sample with the same lithological category as the target sample, but taken from a different well / well section; the meaning of this assignment operation is: the updated feature combination of the target sample retains the original specific causal characteristics. Simultaneously discard the original pseudo-features Introducing pseudo-features of homologous samples .

[0040] By superimposing causal features with the updated pseudo-features, intra-class perturbation samples are generated. ; Based on the constraint of the cross-entropy loss function that the model relies solely on causal features to complete the prediction, the intra-class causal intervention loss is calculated: ; Explanation of parameters in the formula: These are different sample numbers within the same lithological category; These are samples with intra-class perturbations. For intraclass causal intervention loss; Let be the cross-entropy loss function, which is a well-known loss function for classification tasks; For the first The actual lithological label corresponding to each sample.

[0041] (2) Calculation of inter-class causal perturbation and inter-class intervention loss: To enhance the distinguishability of features among different lithologies, reduce the probability of lithological feature confusion, and improve the model's recognition boundary for various strata, this invention adds an inter-class causal perturbation strategy. The core idea of ​​this strategy is: to fix the inherent causal features of the current target sample, discard all interfering pseudo-features carried by the sample itself, and introduce exogenous causal features from other lithological categories to complete feature reconstruction; through cross-fusion of heterogeneous causal features, the model is forced to determine lithology solely based on its own exclusive causal features, weakening the interference of irrelevant features. The specific operation method and perturbation rules are as follows: ; In the formula: This represents the target sample currently awaiting processing. This represents a heterogeneous sample whose lithological category is completely different from that of the target sample; the meaning of this assignment operation is: when updating the sample feature combination, the causal characteristics specific to the target sample are fully preserved. Directly remove original false features Furthermore, causal characteristics of heterogeneous lithological samples were introduced. Fill to the pseudo-feature location.

[0042] By superimposing the original causal features and the external causal features, inter-class perturbation samples are generated: ; Calculate the inter-class causal intervention loss by combining the cross-entropy loss function: ; Explanation of parameters in the formula: Number the different types of samples whose lithology differs from the current sample; These are combined features after inter-class perturbation; This represents the loss due to causal intervention between classes.

[0043] (3) Construction of the global joint loss function: Fixed balancing weights are difficult to adapt to the data offset differences between different wells in multi-well exploration scenarios. This invention sets dynamic adaptive balancing weights, which rely on KL divergence to quantify the degree of feature distribution offset between newly drilled wells and historical wells, and adaptively adjust the proportion of the two types of causal losses. The weight calculation formula is as follows: ; ; Explanation of parameters in the formula: The well-to-well domain distribution divergence; is the KL divergence function, used to measure the degree of difference in characteristic probability distributions between different wells, and is a well-known statistical function; The characteristic probability distribution function; The causal characteristics of newly drilled wells; The cumulative causal characteristics of historically trained drilling; This refers to the divergence scaling hyperparameter; The weighting is dynamically adaptive. The greater the geological differences between wells, The value approaches 1, indicating that the model emphasizes intra-class intervention and weakens irrelevant interference between wells; the higher the geological similarity between wells, the better. Approaching 0, the model focuses on inter-class intervention to enhance the extraction of inherent lithological features.

[0044] Based on the dynamic weight fusion of the three types of losses, a global joint optimization loss function is obtained: ; Explanation of parameters in the formula: To optimize the total loss during the final backpropagation of the model; The mask orthogonal loss balancing hyperparameter is used to adaptively adjust the weight of orthogonal constraints in the total loss, and is a known adjustable hyperparameter; the total loss simultaneously integrates the intra-class intervention loss, inter-class intervention loss, and mask orthogonal regularization loss, taking into account feature robustness, class discriminability, and feature decoupling purity.

[0045] (4) Domain Increment Hierarchical Iterative Training: Following the drilling access sequence, model training is completed sequentially for each well. The overall training logic does not require repeatedly storing historical drilling samples; only the iteratively updated network weight parameters are saved, resulting in low memory overhead and adaptability to engineering operation scenarios where new exploration wells are continuously added to oil and gas work areas. After completing multiple rounds of incremental domain iterative training, the model network parameters are solidified, enabling automatic lithology prediction for newly added unknown wells within the work area. The complete implementation process is as follows: Initial domain training: for the first incremental domain (the first historical well). The network parameters of the dual-channel encoder, orthogonal mask generation module, and lithology classifier are randomly initialized. The parameters are iteratively updated using the global joint loss function, and the optimal basic model corresponding to the incremental domain is saved.

[0046] Incremental domain iterative training: For newly added wells that are subsequently connected, i.e., the incremental domain satisfies... When training, the optimal network parameters from the previous incremental domain training are loaded as the initial parameters for this training. The general lithological features learned from historical drilling are reused, eliminating the need to reinitialize all parameters and reducing the model training cost.

[0047] Network parameter update: Based on the training set data corresponding to the current incremental domain, using the aforementioned global joint loss... Backpropagation is used to synchronously update the network parameters of the encoding branch, mask generation module, and classifier; during iterative training, the model accuracy is monitored based on the validation set until the model performance converges, and the optimal model parameters in the current incremental domain are saved.

[0048] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0049] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0050] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0052] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A well logging lithology prediction method based on multi-well incremental learning, characterized in that, The construction and training process of the prediction model used includes: Construct a global incremental dataset with each well as an incremental unit, and preprocess and standardize the logging data of each well. A dual-channel coding structure is constructed, and the logging time-series features of each logging data sample in the global incremental dataset are extracted along the time branch and the channel branch respectively. Based on the formation data fluctuation state, the dual-branch features are adaptively fused to obtain the global time-series representation. Based on global temporal representation, causal feature masks and pseudo-feature masks are generated through a dual-branch independent mapping network. An F-norm-based mask orthogonal regularization loss is introduced to make the two types of masks orthogonal. The masks are used to decouple the global temporal representation into lithological causal features and domain-specific pseudo-features. Intra-class causal perturbation and inter-class causal perturbation are performed on the decoupled lithological causal features and domain-specific pseudo-features, respectively. Intra-class causal intervention loss and inter-class causal intervention loss are constructed based on the lithological classifier. The intra-class causal intervention loss, inter-class causal intervention loss and mask orthogonal regularization loss are fused based on dynamic adaptive weights to construct a global joint loss function. Using a single well as an incremental unit, historical prediction model parameters are loaded sequentially for incremental iterative training, and the current prediction model parameters are saved. This process is repeated for multiple wells to complete incremental learning.

2. The well logging lithology prediction method based on multi-well incremental learning according to claim 1, characterized in that, The construction of the global incremental dataset with single wells as the incremental unit specifically includes: Each well within the exploration area is divided into an independent incremental domain, and a global incremental dataset is generated. ,in, , This represents the total number of wells drilled. For the first The subset of data corresponding to the wellhead drilling. For the first The first drilling in the well well logging data sample, For the number of logging data channels, for lithological labels For the first Total number of logging data samples in wells; It is the set of real numbers.

3. The well logging lithology prediction method based on multi-well incremental learning according to claim 1, characterized in that, The construction of the dual-channel encoding structure, which extracts the logging time-series features of each logging data sample in the global incremental dataset along the time branch and channel branch respectively, specifically includes: The dual-channel encoding structure includes a time encoder and a channel encoder; well logging data samples Perform sliding window segmentation to obtain time series samples. The time encoder uses time-series samples For input, a two-layer one-dimensional convolutional layer is used to extract local temporal correlation features along the well depth direction layer by layer. After batch normalization, ReLU activation, and average pooling, the temporal evolution representation is output. The channel encoder takes the transposed temporal sample as input, uses a two-layer one-dimensional convolutional layer to complete feature interaction in the channel dimension, and after batch normalization and ReLU activation, outputs a multi-parameter coupled representation. Modeling the intrinsic physical relationships between multiple parameters: ； ； The temporal evolution representation of the output of the time branch; Multi-parameter coupling characterization of channel branch output; This is a matrix transpose operation; These are the learnable network parameters corresponding to the time encoder and the channel encoder, respectively; the logging time series features include time series evolution representation and multi-parameter coupling representation.

4. The well logging lithology prediction method based on multi-well incremental learning according to claim 1, characterized in that, The adaptive fusion of dual-branch features based on the fluctuation state of stratigraphic data to obtain a global time series representation specifically includes: Dynamically calculate fusion weights based on the temporal variance of a single batch of samples: ； For adaptive fusion weights; For activation functions; This is the variance calculation function; Representing the Inside the well, the first Time-series data corresponding to each logging data channel; Number the wells; This represents the total number of well logging data channels. By fusing features from the time branch and channel branch outputs using adaptive fusion weights, an enhanced global temporal representation is obtained. And connected to the lithology classifier Preliminary lithological prediction completed: ； ； The temporal evolution representation of the output of the time branch; This represents the multi-parameter coupling of the channel branch output. For the first All learnable network weights and bias parameters corresponding to the lithology classifier for each well; This is for preliminary prediction of lithological results.

5. The well logging lithology prediction method based on multi-well incremental learning according to claim 1, characterized in that, The method, based on global temporal representation, generates causal feature masks and pseudo-feature masks through a dual-branch independent mapping network, and introduces an F-norm-based mask orthogonal regularization loss to make the two types of masks orthogonal. Specifically, this includes: ； ； These are two sets of single-layer fully connected neural networks with identical structures and independent parameters. As a causal feature mask; This is a pseudo-feature mask; The normalized activation function; Introducing an F-norm-based mask orthogonal regularization loss Quantify the overlap between causal feature masks and pseudo-feature masks, by Backpropagation minimizes the matrix inner product, achieving explicit orthogonality constraints: ； This represents the matrix transpose operation; It is the F-norm.

6. The well logging lithology prediction method based on multi-well incremental learning according to claim 1, characterized in that, The method of decoupling the global temporal representation into lithological causal features and domain-specific pseudo-features using masks specifically includes: It employs element-wise multiplication and utilizes orthogonal causal feature masks. and pseudo-feature mask Global temporal representation Feature decoupling and separation were performed to isolate lithological causal characteristics. Domain-specific pseudo-features : ； ； This is an element-wise matrix multiplication operation.

7. The well logging lithology prediction method based on multi-well incremental learning according to claim 1, characterized in that, The process of performing intra-class causal perturbations and inter-class causal perturbations on the decoupled lithological causal features and domain-specific pseudo-features, respectively, specifically includes: Intraclass causal perturbation: For any target logging data sample, other logging data samples with the same lithology label as the target logging data sample are selected from the global incremental dataset. The domain-specific pseudo-features of the selected logging data samples are extracted and replaced with the domain-specific pseudo-features of the target logging data sample, while retaining the lithological causal features of the target logging data sample itself. Superimposed lithological causal characteristics With the updated domain-specific pseudo-features Generate intra-class perturbation samples : ； An index for the target well logging data sample; An index representing other well log samples whose lithology labels match the target well log sample; Inter-class causal perturbation: For any target well logging data sample, and with the introduction of lithological causal features of well logging data samples with other lithological labels, the domain-specific pseudo-features of the target well logging data sample are replaced, while preserving the lithological causal features of the target well logging data sample itself. Superimposed lithological causal characteristics Based on the causal characteristics of exogenous lithology, inter-class perturbation samples are generated. : ； An index representing a logging data sample whose lithology label differs from the target logging data sample. Lithological causal characteristics of logging data samples that differ from the lithological label of the target logging data sample.

8. The well logging lithology prediction method based on multi-well incremental learning according to claim 7, characterized in that, The global joint loss function, constructed by fusing intra-class causal intervention loss, inter-class causal intervention loss, and mask orthogonality regularization loss based on dynamic adaptive weights, specifically includes: Calculate intraclass causal intervention loss : ； For the i-th target logging data sample, there is an intra-class disturbance sample; The cross-entropy loss function; The true lithology label for the i-th target logging data sample; As a lithology classifier, These are all the learnable network weights and bias parameters corresponding to the lithology classifier; Calculate inter-class causal intervention loss : ； For the inter-class perturbation sample of the i-th target logging data sample; Intra-class causal intervention loss based on dynamic weight fusion Inter-class causal intervention loss Mask orthogonal regularization loss The global joint loss function is obtained. : ； The mask orthogonal regularization loss balancing hyperparameter, It is a dynamic adaptive weight, calculated based on KL divergence, used to quantify the distribution offset between the current drilling lithological causal characteristics and the historical drilling lithological causal characteristics.

9. The well logging lithology prediction method based on multi-well incremental learning according to claim 8, characterized in that, The dynamic adaptive weights are calculated as follows: ； ； The well-to-well domain distribution divergence; Let KL divergence function be used. The characteristic probability distribution function; The lithological causal characteristics of newly drilled wells; The cumulative lithological causal characteristics of historically trained wells; This refers to the divergence scaling hyperparameter; For dynamic adaptive weights, This is the activation function.

10. The well logging lithology prediction method based on multi-well incremental learning according to claim 1, characterized in that, The process of incrementally training multiple wells by sequentially loading historical prediction model parameters using a single well as an incremental unit, and saving the current prediction model parameters, includes: During the first incremental domain training, the prediction model parameters are randomly initialized, the prediction model parameters are updated based on the global joint loss function, and the optimal prediction model is saved. During subsequent incremental domain training, the optimal prediction model parameters from the previous incremental domain are loaded as initial parameters, and the prediction model parameters are iteratively updated based on the current incremental domain data until the prediction model converges and the optimal model parameters of the current incremental domain are saved. During training, historical drilling and logging data samples are not cached; only the prediction model parameters are saved. The prediction model parameters include those of the dual-channel coding structure, the bi-branch independent mapping network, and the lithology classifier.

Images